| 高性能计算机系统接口通信机制的研究 |
| |
| 引用本文： | 戴梅萼,张灏,周桂林. 高性能计算机系统接口通信机制的研究[J]. 小型微型计算机系统, 2001, 22(3): 278-279 |
| |
| 作者姓名： | 戴梅萼 张灏 周桂林 |
| |
| 作者单位： | 清华大学计算机科学与技术系北京 100084 |
| |
| 摘 要： | 高性能并行计算机技术是当今关键技术之一，由清华大学计算机系设计的高性能网络并行超级计算系统THNPSC－1在设计接口时，通过一个表对通信缓冲区进行维护，这样可以直接将数据从应用空间输入／输出到网络，从而支持真正意义上的0拷贝通信协议。通过提供用户层网络访问接口，通信与计算重叠等手段，该接口取得了很高的通信性能。其中实际测量的应用程序之间16字节信息的传递延迟不到7us，最大持续通信带宽可达30MBytes/s。
|
| 关 键 词： | 用户层通信 并行计算机系统 接口 通信机制 |
| 文章编号： | 1000-1220(2001)03-0278-02 |
RESEARCH FOR COMMUNICATIONMECHANISM OF A HIGH PERFORMANCE COMPUTER SYSTEM |
| |
| DAI Mei e ZHANG Hao ZHOU Gui ling. RESEARCH FOR COMMUNICATIONMECHANISM OF A HIGH PERFORMANCE COMPUTER SYSTEM[J]. Mini-micro Systems, 2001, 22(3): 278-279 |
| |
| Authors: | DAI Mei e ZHANG Hao ZHOU Gui ling |
| |
| Abstract: | The parallel super computation is a key technology now.The high performance network parallel super computation system for cluster THNPSC 1 designed by the Department of Computer Science and Technology of Tsinghua University safeguards the communication buffer by using a table. It can directly transport message in/out to network, give sufficiently support to truly zero_copy communication protocol implementation. By providing user_level network access interface and overlapping communication with computation,the high communication performance has been got. The delay of 16 bytes message measured at application is less then 7us, and the sustained bandwidth reaches to 30MBytes/s for 2K bytes length message. |
| |
| Keywords: | Workstation clusters Network interface User level communication Overlapping |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
